Orchestrate parallel coding agents through the full research loop with human checkpoints

Sakana AI's Agentic Scientist, the third version of its AI Scientist, wraps coding agents in a research scaffold. Parallel workers each propose a hypothesis with rubrics for what an experiment must show, edit and run code, and share findings and failures through common memory. The best correct program is then ablated, analysed and written up as a LaTeX paper with an evidence map tying claims to logs, checked by audit and reviewer agents. Using an early version, Sakana researchers studied Error Diffusion, a biologically inspired alternative to backpropagation; after they verified results and references, the paper was accepted as an ALIFE 2026 oral. Another run raised exact match retrieval from 52% to 98% on a synthetic test. Human inspection also caught a reward hacked candidate.